Cursor tells Congress it may phase out Chinese open-weight AI models
Updated: 8 hours ago

Cursor maker Anysphere has told two House committees that it plans to explore phasing out open-weight AI models developed in China because of geopolitical risk. The statement appears in a joint investigative report released on November 20, 2026 by the House Committee on Homeland Security, chaired by Andrew Garbarino (R-NY), and the House Select Committee on China, chaired by John Moolenaar (R-MI). The report, titled "China is Trying to Get U.S. Companies Addicted to Chinese AI Models," closes an inquiry that the two chairmen opened on April 29, 2026 with letters to Anysphere chief executive Michael Truell and to Airbnb, and later widened to DoorDash.
The admission carries weight for Cursor users because the company's own flagship coding models sit on a Chinese base. Composer 2, released in March, was built on Moonshot AI's open-weight Kimi K2.5, and Cursor says Composer 2.5, the current version launched on May 18, 2026, is "built on the same open-source checkpoint." Composer 2.5 is one of the models included in the weekly usage pool of every paid Cursor plan, priced at $0.50 per million input tokens and $2.50 per million output tokens when usage goes beyond that pool.
The report gives no timetable and no decision. It records only that Anysphere "plans to explore phasing out PRC-developed open-weight models," citing a staff briefing, and it sets that position against DoorDash and Airbnb, which kept using Chinese models after the government warnings the report describes. Cursor's public release notes, checked on November 22, list no change to Composer or to the Chinese models in its model picker.
··········
WHAT THE HOUSE REPORT SAYS ABOUT ANYSPHERE AND CURSOR.
The Anysphere material in the report is short and built on three kinds of evidence: the Composer 2 technical report of March 26, 2026, which the company produced to the committees, a staff briefing, and Cursor's public security page. The committees write that the link between Composer 2 and a Moonshot AI base model became public only after an independent developer identified the underlying model, a point our earlier coverage of the Composer 2 and Kimi K2.5 controversy traced in March.
According to the report, Anysphere described Chinese open-weight models to the committees as competitive options in their own right, and as no fallback of last resort. The company also described a vendor risk management program and technical controls that govern infrastructure and model access. The committees add that company submissions overall gave considerably less detail about how they evaluated the Chinese developers themselves, and they cite Cursor's security page among those submissions.
The sentence that drew attention comes near the end of the company findings. Anysphere "informed the Committees that it plans to explore phasing out PRC-developed open-weight models because of geopolitical risk." The wording describes an exploration, and the report attaches no date, budget or replacement model to it.
Cursor has a candidate replacement path on record. On its Composer page the company says it is training "a significantly larger model from scratch" together with SpaceXAI, using ten times more total compute on the Colossus 2 cluster. A model trained from scratch would carry no Kimi K2.5 weights, though Cursor has given no release date for it and has not linked it to the House inquiry.
··········
CHINESE-DEVELOPED MODELS CURRENTLY OFFERED IN CURSOR.
Composer is the most visible Chinese-based model in Cursor, but the model picker also lists third-party models from three Chinese developers. Cursor's models and pricing page shows them with the following list prices, as checked on November 22, 2026.
........
Model | Developer | Input | Cache read | Output | In paid plan pool |
Composer 2.5 | Cursor (Kimi K2.5 base) | $0.50 | $0.20 | $2.50 | Yes |
Composer 2.5 Fast | Cursor (Kimi K2.5 base) | $3.00 | $0.50 | $15.00 | Yes |
Kimi K3 | Moonshot AI | $3.00 | $0.30 | $15.00 | No |
Kimi K2.7 Code | Moonshot AI | $0.95 | $0.19 | $4.00 | No |
GLM 5.3 | Z.ai | $1.40 | $0.26 | $4.40 | No |
GLM 5.2 | Z.ai | $1.40 | $0.26 | $4.40 | No |
GLM 5.3 Flash | Z.ai | $0.15 | $0.029 | $0.50 | No |
DeepSeek V4.1 Flash 1017 | DeepSeek | $0.30 | $0.006 | $1.20 | No |
........
Prices are in USD per million tokens. The third-party models sit outside the included pool: on Premium and higher plans they are billed at API list price from prepaid credit, and the Plus and Free plans do not include them. Teams and Enterprise accounts pay an extra Cursor Token Rate of $0.25 per million tokens on eligible third-party models, with an exemption for models whose cache-read price is below $0.10, which covers DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. Cursor says the third-party models run through inference partners it does not name on the model pages.
The report does not say whether a phase-out would cover only Composer's base model, the third-party Chinese models in the picker, or both. The table shows how much of Cursor's low-cost tier depends on them.
··········
WHAT A CURSOR AGENT SESSION COSTS ON THESE MODELS.
To put the price gaps in concrete terms, take an illustrative agent session that reads 2 million input tokens, with 80% of them served from cache, and writes 60,000 output tokens. The arithmetic below uses the list prices above and ignores the Teams surcharge and the included pool, so it shows marginal cost only.
........
Model | Uncached input (400K) | Cached input (1.6M) | Output (60K) | Session cost |
Composer 2.5 | $0.20 | $0.32 | $0.15 | $0.67 |
Composer 2.5 Fast | $1.20 | $0.80 | $0.90 | $2.90 |
Kimi K3 | $1.20 | $0.48 | $0.90 | $2.58 |
Kimi K2.7 Code | $0.38 | $0.30 | $0.24 | $0.92 |
GLM 5.3 | $0.56 | $0.42 | $0.26 | $1.24 |
DeepSeek V4.1 Flash 1017 | $0.12 | $0.01 | $0.07 | $0.20 |
GLM 5.3 Flash | $0.06 | $0.05 | $0.03 | $0.14 |
........
The two cheapest options in the table, DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash, both come from Chinese developers, and standard Composer 2.5 costs about a quarter of Kimi K3 or Composer 2.5 Fast for the same session. Those gaps match the House report's own finding that low price and customizability drive adoption of Chinese models, and they show what Cursor would have to replace if it removed them.
··········
[[ADPLUS_300x250_1]]
WHAT CHANGES FOR CURSOR USERS AND TEAMS NOW.
Nothing changes in the product today. The models above remain listed, Composer 2.5 remains in the included pool, and Cursor has published no migration notice. Individual developers have no action to take until Cursor announces a decision.
Teams that sell to the federal government or operate in critical infrastructure should read the report's recommendations closely, because they describe rules that could reach tools like Cursor if Congress or agencies adopt them. The committees recommend extending the existing exclusion of DeepSeek from federal civilian systems to bar Chinese-developed models and their derivatives from federal systems and from federal contract performance, with narrow waivers. A fine-tuned derivative of Kimi K2.5 would plausibly fall under that language, although the report does not name Composer in its recommendations.
Further recommendations call for NIST and CISA to create a model provenance standard and an AI bill of materials, required for federal contractors and high-risk critical infrastructure, and for a phase-out of Chinese models in high-risk critical infrastructure functions. Under the committees' proposal for the PILLAR Act, State and Local Cybersecurity Grant Program funds could not be spent on Chinese models, while replacement costs would remain eligible. None of these measures is law yet, and the report is a committee product with no binding force.
For procurement teams the practical step is documentation. Organizations that must answer supplier questionnaires about model provenance can already record which Cursor models their developers use, because Cursor publishes the developer of each model and Composer's Kimi K2.5 base on its own pages. Readers comparing Cursor with alternatives can find plan prices in our guide to AI coding agent pricing, and the background on Cursor's recent vendor problems in our report on the Fireworks access key behind Cursor's security incident.
··········
AIRBNB AND DOORDASH: HOW OTHER COMPANIES USE CHINESE MODELS.
The two other companies in the inquiry described heavier and more lasting use of Chinese models than Anysphere did. Airbnb told the committees that it runs Alibaba's Qwen 2.5, Qwen 3 and Qwen 3.5 on infrastructure from U.S.-based service providers, with no paid relationship with a Chinese developer. Its customer AI assistant handles about 80,000 guest and host interactions a day, and a fine-tuned summarization system processes about 250,000 cases and calls a day across six languages. Zhipu's GLM-4.5 is approved for production.
DoorDash reported 25 Chinese model families authorized and deployed in a non-trivial way, which the report defines as at least $5,000 in compute and serving costs. The company named Qwen variants up to Qwen3.6, DeepSeek V4 Pro, Moonshot's Kimi K2.6 and Kimi K3, Zhipu's GLM-5 through GLM-5.2 and MiniMax M2.5, among others. Kimi K2.6 works as a first-pass code reviewer, and on DoorDash's internal DashBench test of 105 historical code changes, Kimi K2.6 paired with Claude Fable 5 scored 75.3% weighted F1, the best result recorded. DoorDash said most uses could be switched almost immediately, while replacing one model embedded in a larger system could take up to six months. It also blocks connections to Chinese AI services from corporate devices and model hosting environments.
........
Company | Chinese models named in the report | Scale disclosed | Stated plan |
Anysphere (Cursor) | Kimi K2.5 as Composer base | No usage figures given | Explore phasing out |
Airbnb | Qwen 2.5, Qwen 3, Qwen 3.5, GLM-4.5 | ~80,000 chats and ~250,000 summaries a day | Continued use |
DoorDash | 25 families incl. Qwen, DeepSeek, Kimi, GLM, MiniMax | 25 families above $5,000 each | Continued use |
........
··········
THE COMMITTEES' FINDINGS ON CHINESE OPEN-WEIGHT MODELS.
The report's five findings frame the Cursor statement. The committees write that Chinese open-weight models accounted for 41% of Hugging Face downloads over the prior year and that spending data understates their use, citing one gateway where DeepSeek handled 17% of tokens but about 1% of spending. They note that the highest-ranked U.S. open-weight model placed tenth on the Artificial Analysis index, behind six Chinese models.
The committees also allege that six Chinese developers ran industrial-scale distillation campaigns against American models, likely with government awareness, and that U.S. data-labeling firms earn about $500 million a year from Chinese labs. On security, the report describes testing of Zhipu's GLM-5.3, whose refusal rate on harmful requests fell from about 95% to between 3% and 6% after researchers modified its open weights, at a cost of about 2,200 GPU hours or roughly $4,400. NIST found the model about four months behind the American frontier on an aggregate of government cyber benchmarks. GLM-5.3 is listed in Cursor's model picker as a third-party model.
The twelve recommendations are grouped under five goals: building an American open-weight alternative, knowing which models are running, imposing costs on illicit extraction, setting durable rules for closed and open models, and working with allies. The first recommendation is a national American open-weight AI initiative with public compute and federal procurement demand, which the committees present as the answer to the price gap that drew companies to Chinese models.
··········
CURSOR AND CHINESE MODELS: CONFIRMED FACTS AND OPEN QUESTIONS.
The confirmed facts are narrow. A House report dated November 20, 2026 states that Anysphere told the committees it plans to explore phasing out Chinese-developed open-weight models because of geopolitical risk. Cursor's own pages confirm that Composer 2.5, which is included in every paid plan's usage pool, is built on Moonshot's Kimi K2.5 checkpoint, and that the picker offers six more models from Moonshot, Z.ai and DeepSeek.
The open questions are larger. Cursor has published no statement, timeline or scope for the exploration, and the report does not say whether it would cover Composer, the third-party Chinese models, or both. The from-scratch model Cursor is training with SpaceXAI is the obvious successor to a Kimi-based Composer, but its release date and price are unknown, and so is whether it could match Composer 2.5's $0.50 and $2.50 list prices. For federal contractors and critical infrastructure operators, the report's language on derivatives and provenance is the part to track, since it is the route by which a Kimi-based Composer could become a compliance question for Cursor customers.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




